“Apple Watch is reliable about when and how long you slept, reasonable on Core and REM, and weakest on deep sleep, so treat stage minutes as estimates with a half-hour margin and judge trends over weeks rather than single nights.”
The short answer
Apple Watch is good at telling sleep from wake and only moderately good at telling the stages apart. In Apple's own validation against polysomnography (PSG), the laboratory gold standard, the current algorithm recognizes 96.8% of sleep epochs and 78.9% of wake epochs, and agrees with human scorers on the four stages with a kappa of 0.681. Independent laboratories that tested the Watch without Apple's involvement report the same strength on sleep versus wake and the same weak spot: deep sleep, which the Watch finds in only about half of the 30-second windows where the EEG sees it23.
Kappa is a chance-corrected agreement score: 1.0 is perfect, 0 is a coin flip. For context, trained technologists scoring the same PSG recording agree with each other at a kappa of about 0.7610. Over a whole night the Watch is therefore not far from a second human opinion, but its errors cluster in deep sleep and in quiet, motionless wake.
Everything below refers to Apple's own sleep stages feature, introduced with watchOS 9 in 2022 and retrained for watchOS 26 in September 20251. As of watchOS 27 and the Apple Watch Series 12, released in September 2026, Apple has not published a newer validation document, so the October 2025 figures are the current ones115.
How Apple Watch estimates sleep stages
A sleep lab scores sleep from brain waves, eye movements and muscle tone. Apple Watch has none of those sensors. Its staging classifier works from the 3-axis accelerometer alone, reading gross movements and the tiny wrist oscillations caused by breathing, and assigns every 30-second epoch to one of four labels: Awake, REM, Core or Deep1. Heart rate, wrist temperature and blood oxygen are recorded overnight for other features; the stage classifier itself is accelerometer-based1.
The labels map onto the clinical ones with one simplification. Deep is N3, REM is REM, and Core merges N1 and N2 because the Watch cannot separate them. Apple chose the word Core rather than "light" because N2 normally makes up more than half of a night and is not a lesser kind of sleep1.
The algorithm was trained on 1,171 nights from 858 people and tested on a separate 299 nights from 166 people, including participants with sleep apnea and periodic limb movements1. In 2025 Apple retrained it with foundation models built on Apple Heart and Movement Study data and evaluated the new version on the same validation set1. Since watchOS 11 the Watch also tracks sleep outside a schedule, but it writes stages only for sessions longer than three hours; shorter naps get a plain asleep-or-awake record1.
Sleep versus wake: the strong part
Two numbers matter here. Sensitivity is the share of true sleep epochs the Watch labels as sleep; specificity is the share of true wake epochs it labels as wake. Apple reports medians of 97.9% and 75.0% for the 2022 algorithm, and 96.8% and 78.9% for the 2025 update1. Traditional research actigraphy typically reaches 90 to 97% sensitivity but 50% specificity or less1.
Independent studies confirm the sensitivity and are harsher on specificity. In a 2024 single-night study of 35 healthy adults, the Apple Watch Series 8 detected 97% of sleep epochs and agreed with PSG on sleep versus wake in 93% of all epochs3. In a 2025 study of 62 adults referred for suspected sleep apnea, sensitivity was 96% but specificity fell to 52%, so almost half of the night's wake epochs were recorded as sleep2.
The practical consequence is a consistent bias: the Watch tends to report slightly more sleep than you got. Total sleep time was overestimated by 8 minutes in the healthy group and 20 minutes in the apnea group, and wake after sleep onset underestimated by 10 and 21 minutes23. A 2025 meta-analysis of 24 wrist-tracker studies, Apple Watch among them, put the pooled discrepancy against PSG at about 17 minutes of total sleep time and 13 minutes of wake after sleep onset7. For a seven-hour night that is an error of roughly 3 to 5%.
Stages: Core is solid, REM is decent, Deep is the weak spot
Apple's confusion matrix for the 2025 algorithm shows how often each true stage was labeled correctly: Core 82%, REM 82%, Awake 79% and Deep 68%1. The 2022 version scored 83%, 78%, 70% and 62% on the same four, so the update helped most with wake and deep sleep, and the median four-stage kappa rose from 0.63 to 0.681.
The same pattern appears outside Apple. In the 2024 healthy-adult study, the Series 8 found 86% of light-sleep epochs and 83% of REM epochs but only 51% of deep-sleep epochs, underestimating deep sleep by 43 minutes and overestimating light sleep by 45 minutes; four-stage agreement with PSG was 75% (kappa 0.60)3. In the 2025 six-device study the Watch again led on light (83%) and REM (69%) but reached only 51% on deep, underestimating it by 25 minutes. Its kappa of 0.53 was still the best of the six wrist devices, ahead of Fitbit Sense (0.42), Whoop 4.0 (0.37) and Garmin Vivosmart 4 (0.21)2.
One nuance works in the Watch's favor. When it does label an epoch Deep, it is usually right: precision for deep sleep was 88% in the healthy-adult study3. The problem is the deep sleep it misses, not the deep sleep it invents. A night showing 40 minutes of deep sleep probably contained at least that much, and possibly a good deal more. For what this stage does and how much of it is normal, see deep sleep basics.
Not every study agrees on the direction of the error. A 2026 home-PSG study of healthy adults wearing a Series 7 found the opposite bias, light sleep underestimated by 74 minutes and REM overestimated by 30, although the Watch still ranked first of the three devices tested5. Different watch generations, software versions and bedrooms produce different errors, so there is no fixed correction to apply to your own numbers.
Where the Watch goes wrong, and why
Quiet wake looks like Core
Lying still with your eyes open produces almost no motion. In Apple's validation, 27% of true wake epochs were labeled Core by the 2022 algorithm; the 2025 update, aimed specifically at "quiet wake", cut that to 17%1. This is why the Watch underestimates time awake, and why specificity can drop to 52% when sleep is fragmented2.
Deep sleep looks like Core too
The most common error in Apple's matrix is true deep sleep labeled Core: 38% of deep epochs in 2022, 32% in 20251. Deep and Core are both still, both have regular breathing, and the feature that separates them, slow brain waves, is invisible to a wrist sensor. Research groups that add heart rate to raw Apple Watch accelerometer data reach about 71% four-class accuracy against an EEG headband, so the ceiling for wrist signals is not far above where Apple already is9.
Fragmented sleep is harder
Apple's own analysis found that frequent stage transitions and a higher apnea-hypopnea index were the two factors that most reduced agreement, while the watch model made no difference1. In a separate cohort of 236 sleep-clinic patients, kappa was 0.55 for the 2022 algorithm and 0.66 for the 2025 one1. Independent clinic studies are tougher. A 2023 multicenter Korean study of 75 patients with a mean apnea-hypopnea index of 18, run by a sleep-technology company whose own app came out on top, recorded an Apple Watch 8 kappa of 0.30, with deep sleep found in 41% of epochs and REM in 43%4. A 2026 methods paper built on Apple Watch data from people with sleep apnea adds a warning: long stretches of "easy" wake, such as time in bed before falling asleep, inflate reported accuracy, so figures from healthy volunteers do not transfer to disturbed sleep8.
Fit and session length
What changed with watchOS 26, watchOS 27 and the newest watches
watchOS 26, released in September 2025, brought the retrained accelerometer model described above and introduced the Sleep Score114. Apple says the score's algorithms were developed and tested on more than 5 million nights from the Apple Heart and Movement Study and informed by guidance from the American Academy of Sleep Medicine, the National Sleep Foundation and the World Sleep Society14. A later 26.2 update refined the score classifications; the current bands are Very Low (0 to 40), Low (41 to 60), OK (61 to 80), High (81 to 95) and Very High (96 to 100)1213.
Apple Watch Series 11 added no sleep-specific sensor; the Sleep Score runs on Series 6 or later, SE (2nd generation) or later and every Ultra with iOS 2614. Series 12 and Ultra 4, announced in September 2026, bring a redesigned Health Sensing System that reads heart rate every five seconds all day and measures HRV up to 24 times more often, adding a Recovery HRV reading to overnight vitals and a 0 to 10 Readiness score that takes the sleep score as one input15. None of this changes how stages are scored: Apple's staging document still describes an accelerometer-only classifier, its validation found no difference in kappa between watch models, and no new stage-accuracy figures exist for watchOS 27 or the new hardware115.
The numbers, study by study
Stage percentages are the share of true epochs of that stage the Watch labeled correctly; biases are Watch minus PSG for the whole night.
| Study (year) | Who was tested | Sleep vs wake | Stages | Main bias |
|---|---|---|---|---|
| Apple validation, 2025 algorithm (watchOS 26)1 | 166 adults, 299 nights, in-lab and home PSG; some with sleep apnea | Sensitivity 96.8%, specificity 78.9% | Core 82%, REM 82%, Awake 79%, Deep 68%; kappa 0.68 | Deep labeled Core in 32% of deep epochs |
| Apple validation, 2022 algorithm (watchOS 9)1 | Same validation set; separate clinic cohort of 236 patients | Sensitivity 97.9%, specificity 75.0% | Core 83%, REM 78%, Awake 70%, Deep 62%; kappa 0.63 (0.55 in clinic cohort) | Deep labeled Core in 38% of deep epochs |
| Robbins et al., Sensors, 20243 | 35 healthy adults aged 20 to 50, one lab night, Series 8 | Sensitivity 97%, 93% epoch agreement | Light 86%, REM 83%, Deep 51%; kappa 0.60 | Deep −43 min, light +45 min, WASO −10 min |
| Schyvens et al., Sleep Advances, 20252 | 62 adults referred for suspected apnea (20 wore a Series 8), one lab night | Sensitivity 96%, specificity 52% | Light 83%, REM 69%, Deep 51%; kappa 0.53 (best of six wrist devices) | Total sleep +20 min, WASO −21 min, deep −25 min |
| Lee et al., JMIR mHealth and uHealth, 20234 | 75 sleep-clinic patients, mean apnea-hypopnea index 18, Apple Watch 8 | Four-stage accuracy 56% | Awake 45%, light 66%, deep 41%, REM 43%; kappa 0.30 | Wake epochs often labeled light sleep |
| Alhejaili et al., Sleep Medicine, 20265 | 19 healthy adults aged 18 to 45, home PSG plus 7 days of actigraphy, Series 7 | Agreement described as weak, wide limits | Ranked first of three devices tested | Light −74 min, REM +30 min |
| Lee et al., Journal of Clinical Sleep Medicine, 2025 (meta-analysis)7 | 24 studies, 798 participants, 12 wrist devices including Apple Watch | Pooled total sleep time differs by about 17 min, WASO by about 13 min | Not pooled | Sleep efficiency about 4.7 points |
| Lambe et al., npj Digital Medicine, 2026 (living review)6 | 82 Apple Watch validation studies across 14 metrics | Sleep accuracy rated moderate | Not pooled | Heart rate bias −0.3 bpm, for comparison |
What this means for reading your own data
Trust the frame, question the fill. Bedtime, wake time, total sleep and the regularity of your schedule come from the part of the algorithm with the strongest evidence. They are also what Apple itself scores: the Sleep Score is built from 50 points for duration, 30 for bedtime consistency over the last 13 nights and 20 for interruptions, with no points for stage minutes1112. If you track one thing, track consistency.
Read stage minutes with a margin. Across the studies above, deep sleep was off by 25 to 43 minutes a night and light sleep by 45 to 74 minutes, in one direction or the other235. A single night of "only 35 minutes of deep sleep" sits inside that error bar and tells you very little. Over 30 nights the random part of the error largely cancels, which is why trends over weeks mean something even when single nights do not. If you check the graph every morning, read how to stop over-analyzing Apple Watch sleep data.
Know when the numbers are least reliable. If you snore heavily, wake often, have restless legs or lie awake for long stretches, you are in the population where every study found the weakest agreement124. The Watch is still useful for timing and duration, but a persistent complaint deserves a clinician and a real sleep study rather than a better wearable. Apple itself states that its sleep stages are not intended for clinical use1.
Choose the form factor for the question you have. In the 2024 head-to-head, the Oura Ring Gen3 detected 80% of deep-sleep epochs against the Watch's 51%, while the two were level on sleep versus wake3; our comparison of watches, rings and bedside sensors covers the trade-offs. Moon follows the same reading of the evidence: it imports the stages Apple Watch writes to Apple Health, but leads with duration, timing and consistency and shows stage changes as multi-week trends rather than a nightly verdict.
FAQ
- Is Apple Watch deep sleep accurate?
- Partly. Apple's current algorithm recognizes 68% of deep-sleep epochs; independent labs found about 51%123. When the Watch labels an epoch Deep it is right about 88% of the time, so it under-reports deep sleep rather than inventing it3. Expect the nightly figure to run low by 25 to 43 minutes and compare weeks, not nights23.
- How accurate is Apple Watch REM sleep?
- REM is the second-best stage. Apple reports 82% of REM epochs labeled correctly, and healthy-adult testing found 83%13. In people with suspected sleep apnea the figure fell to 69%, and in a sleep-clinic cohort to 43%24. Most REM errors are REM epochs labeled Core, about 16% in Apple's validation1.
- Is Apple Watch as accurate as a sleep study?
- No. Polysomnography scores sleep from EEG, eye movements and muscle tone; the Watch infers it from wrist motion. Its four-stage kappa of 0.53 to 0.68 is below the 0.76 agreement between two trained human scorers, and it cannot diagnose a sleep disorder1210. Its advantage runs the other way: hundreds of ordinary nights at home instead of one unusual night in a lab.
- Does a newer Apple Watch track sleep stages better?
- Apple's validation found no difference in kappa between watch models; the gains have come from software, chiefly the watchOS 26 retraining that lifted wake detection from 70% to 79%1. The Series 12's new sensors target heart rate and HRV rather than staging, and Apple has published no stage-accuracy data for it15. Keeping watchOS current matters more than the model year.
- Why does my Watch show almost no awake time when I remember waking up?
- Lying still while awake looks like Core sleep to an accelerometer. Apple's algorithm catches 79% of wake epochs in its validation set, and testing in fragmented sleepers found as little as 52%12. The Watch records fewer brief awakenings than PSG, so your real wake time is probably higher than the graph suggests.
Moon reads the sleep stages Apple Watch writes to Apple Health and turns them into weekly trends, so one odd night never becomes a verdict on your sleep. Download Moon for iOS.
